1
$\begingroup$

Suppose that I have a Markov chain that has absorbing states. Since there are absorbing states, lets group the Markov matrix into four blocks: the submatrix all states in the absorbing region(s) $A$, the submatrix of all states that are not in an absorbing region $N$, the transition values from Non-absorbing to absorbing $T$, and then a $0$ block as you cannot move from absorbing to non-absorbing. In essence this means that our Markov matrix $$ M = \begin{bmatrix} N & 0 \\ T & A \end{bmatrix} $$

Note that I have it set up so that the columns of $M$ add to $1$, just so that I do left multiplication of vectors rather than right.

For the equilibrium solution we are looking for a vector $\vec{X}$ such that $$ M\vec{X} = \vec{X} $$ and $$ \sum_i X_i = 1 $$

Now this is a fairly standard proceedure, as you just find the eigenvector corresponding to an eigenvalue of $1$, however I want to show that the equilibrium probabilities for the states in $N$ are always $0$. $$ N\vec{x} = \vec{x} \implies \vec{x} = 0 $$

Here we know that $N$ must contain no trapping regions (not sure how to formally describe this) and at least one column of $N$ must sum to less than $1$.

If I could show that $(N-I)$ was invertible (or that the null space was trivial) or that the magnitude of its largest eigenvalue was $<1$ then that would be sufficient to prove.

I'm trying to do this myself so any hints or pointers to potentially helpful theorems would be very appreciated!

$\endgroup$
1
  • 1
    $\begingroup$ you need a hypothesis that there is a path each state in N to A. For an algebraic approach, consider $M^k$ for $k$ large enough and bound the largest eigenvalue of the block $N^k$... $\endgroup$ Aug 21, 2020 at 23:12

2 Answers 2

2
+50
$\begingroup$

Say we break down the vector $\vec x$ into two parts: $\vec y$ for the non-absorbing states, and $\vec z$ for the absorbing states. Then $M\vec x = \vec x$ tells us that $$ \begin{cases} N \vec y = \vec y \\ T \vec y + A \vec z = \vec z \end{cases} $$ as well us $\sum_i y_i + \sum_j z_j = 1$.

Usually, for the absorbing states, we take $A = I$: once you're in an absorbing state, you stay put. Then $A \vec z = \vec z$, leading us to $T \vec y = \vec 0$.

But even if you don't make this assumption, we can deduce $T \vec y = \vec 0$. Let $\vec 1$ be the all-$1$ vector (of the same dimension as $\vec z$); from $T \vec y + A \vec z = \vec z$, we get $\vec 1^{\mathsf T} T \vec y + \vec1^{\mathsf T}\!A \vec z = \vec1^{\mathsf T}\vec z$. Because the columns of $A$ add up to $1$, we must have $\vec1^{\mathsf T}\!A = \vec1^{\mathsf T}$, so we get $$ \vec1^{\mathsf T} T \vec y + \vec1^{\mathsf T} \vec z = \vec1^{\mathsf T} \vec z \implies \vec1^{\mathsf T} T \vec y = 0. $$ In other words, the components of $T \vec y$ sum to $0$; however, since they're nonnegative, this can only happen if $T \vec y = \vec 0$.


How do we use $T\vec y= \vec 0$?

Look at the $i^{\text{th}}$ component of this product: it says $t_{i1} y_1 + t_{i2} y_2 + \dots + t_{ik} y_k = 0$. Here, every $t_{ij}$ and every $y_j$ is nonnegative. So the only way for the sum to be $0$ is that whenever $t_{ij} > 0$, $y_j$ must be $0$. So all states with a transition to an absorbing state have a limiting probability of $0$.

Next, whenever we deduce $y_j=0$, knowing that $N\vec y = \vec y$ tells us that $(N\vec y)_j = 0$, or $n_{j1} y_1 + n_{j2} y_2 + \dots + n_{jk} y_k = 0$. Here, also, every term is nonnegative; whenever $n_{j\ell} > 0$, $y_\ell$ must be $0$ by the same logic. So all non-absorbing states with a transition to such a state $j$ (a state $j$ which has a transition to an absorbing state) must also have a limiting probability of $0$. To rephrase: all non-absorbing states with a $2$-step path to an absorbing state must have a limiting probability of $0$.

From here, we can prove that all non-absorbing states with a path to an absorbing state must have a limiting probability of $0$, by induction on the length of the path. If we assume that from every non-absorbing state, there's a path to an absorbing state, then we can conclude that $\vec y = \vec 0$.

$\endgroup$
2
$\begingroup$

the standard form in probability is for the matrix to be row stochastic, so I work on the transpose

$M^T = \begin{bmatrix} N^T & * \\ \mathbf 0 & A^T \end{bmatrix}$

you need a hypothesis that each state in $N$ has a path with positive probability to that in $A$, otherwise what you're looking for would not be true, e.g.
$N^T = \begin{bmatrix} 0&1&\mathbf 0\\ 1&0&\mathbf 0\\ 0&0&(N')^T\end{bmatrix}$
would violate what you are trying to prove since state 1 has only a path to state 2 which only has a path to state 1.

so I assume that each state in $N$ has a path to (at least) one state in $A$. A standard exercise for $m$ state chains -- if there is a path from (i) to (j) then it takes at most $m$ iterations for that path to be realized with positive probability (either direct combinatoric proof or for an algebraic proof: apply Cayley Hamilton).

Blocked multiplication tells us
$\big(M^T\big)^m = \begin{bmatrix} (N^T)^m & * \\ \mathbf 0 & (A^T)^m \end{bmatrix}$
and by our assumption the $*$ cells contain a positive entry in each row.

Now since $M^T$ is row stochastic we have
$M^T\mathbf 1_m = \mathbf 1_m\implies (M^T)^m\mathbf 1_m = \mathbf 1_m$
and if we subtract out the positive components in each row of the $*$ cells we see this means that the sums across the rows of $(N^T)^m$ are strictly less than one. That is
$(N^T)^m\mathbf 1 \lt \mathbf 1$ (where the inequality holds component wise.)

Direct application of Gerschgorin discs tells us that $\sigma\big((N^T)^m\big)\lt 1$. This implies $\sigma\big((N^T)\big)\lt 1$, or equivalently since a matrix and its transpose have the same eigenvalues:
$\sigma\big(N\big)\lt 1\implies N\vec{x} = \vec{x} \implies \vec{x} = 0$

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .